AI 客服编造「每订阅限一台设备」政策：用户退订，联合创始人三小时后澄清An AI support agent fabricated a one-device-per-subscription policy: users cancelled, and a co-founder corrected it within three hours

After being signed out across devices, a user asked Cursor support for help and the AI agent confidently stated a core security policy, one device per subscription, that does not exist. The support channel had neither fact-checking nor an escalation path.

Users cancelled and publicised it; a co-founder corrected the record within three hours, acknowledging a front-line AI support hallucination.

The same pattern as INC-008 and INC-010: what was fabricated was a policy, not small talk, and policy claims carry force because users act on them directly. The support channel mistook fluency for reliability and left no check or escalation path for high-risk assertions.

Policy assertions in support answers must be grounded in the policy base; when there is no basis, the agent must say so and hand over to a human. The more real it sounds, the more it needs checking first.

Ground policy claims: support channels verify policy, price and permission assertions against the base; no basis, no answer.

Escalate when unsure: say so and hand over to a human rather than filling the gap fluently.

Sample and alert: audit support conversations and alert on high-risk assertions, with a fast correction path.